Answers are grounded in this paper and cite their sources.
Demonstrates improved detection of bowel obstruction in radiography, indicating potential for better diagnosis.
Detecting and localizing bowel obstruction on upright abdominal radiograph is a challenging task, and ground truth typically relies on imaging examinations, endoscopy or surgical notes within a short time window. The lack of ground truth leads to poor performance on supervised learning. To address this issue, self-supervised learning (SSL) is used to determine the presence and site of bowel obstruction using few labeled samples. A total of 1500 labeled and 1922 unlabeled radiographs were categorized based on the ground truth. The SSL model was firstly trained on the unlabeled radiographs, and then fine-tuned on the labeled radiographs. ResNet50 and VGG16 were used for the embedded base encoders, whose weights and parameters were adjusted during training process. Furthermore, it was tested on an external dataset, compared with supervised learning models and human interpreters. Finally, the t-SNE and Grad-CAM were used to visualize the model’s interpretation. The proposed SSL models showed superior classification performance (VGG16-backboned: accuracy of 0.846, MCC of 0.873, Kappa of 0.770; ResNet50-backboned: 0.833, 0.860, 0.750), outperformed supervised models (Mean: 0.750, 0.792, 0.625) and a senior radiologist (0.748, 0.791, 0.620). Significant statistical differences were observed between VGG16-backboned SSL model and supervised models, as well as human interpreters ( P < 0.05). Additionally, the SSL model exhibited better cluster performance through t-SNE and demonstrated outstanding lesion recognition ability through Grad-CAM. The proposed SSL model exhibited outstanding performance in determining the presence and site of bowel obstruction on upright abdominal radiography based on unlabeled images and limited ground truth, indicating the potential of SSL for computer-aided diagnosis of diseases that are difficult and time-consuming for labeling.
No takes yet. Share an insight, caveat, or question.
A 2026 study studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: